[overbit]

Summary

Add native local System One inference to OpenCode V2's fast-JEV compaction plugin.

• New backend: "systemone" for LM Studio 0.4.26+, Ollama 0.35+, and compatible POST /v1/systemone servers.
• One shared state and multiple noul retention decisions per request; split by 64 questions and a conservative 60 KiB UTF-8 JSON limit.
• Configurable server URL, required local model identifier, and optional local/proxy bearer token. Never forwards the hosted TYPESAFE_API_KEY.
• Conservative System One state/request budgets, with a smaller quick-start budget for Ollama tev1:0.8b.
• Preserves existing hosted TypeSafe and legacy LM Studio chat-completions integrations.
• Tests for LM Studio/Ollama URL resolution, batching, parsing, missing/out-of-range probabilities, partial failures, no-key access, authentication, fail-open behavior, and the V2 compaction hook.
• Updated README and GitHub Pages quick starts.

Usage

LM Studio

{
  "plugins": [{
    "package": "github:overbit/fast-jev-compaction-opencode",
    "options": {
      "backend": "systemone",
      "localModel": "YOUR_LOADED_DECISION_MODEL_ID"
    }
  }]
}

Ollama

ollama pull tev1:0.8b

{
  "plugins": [{
    "package": "github:overbit/fast-jev-compaction-opencode",
    "options": {
      "backend": "systemone",
      "localBaseUrl": "http://127.0.0.1:11434/v1",
      "localModel": "tev1:0.8b",
      "maxStateTokens": 1200,
      "maxRequestTokens": 1800
    }
  }]
}

Verification

• npm run typecheck
• npm test (including plugin and static site checks)
• npm run compile
• Packaged server/TUI plugin load smoke test
• GitHub Actions CI green on the PR head
• Live inference against a user's running LM Studio/Ollama (CI does not have either model server)

This preserves the V2 limitation that the hook can return a deterministic summary string but not replace structured messages. No third-party dependency added.

Original design: tamaratran/fast-jev-compaction.